Notion AI Automations vs Sup AI

“The primitive here is: a visual workflow engine that injects LLM steps between database triggers and HTTP calls — basically Zapier with an AI node, living inside your wiki. The DX bet is that no-code is the right abstraction layer, which means the moment of truth is 'can I actually call my API with a structured payload and handle errors?' — and based on the blog post, there's no answer to that. There's no repo, no webhook schema docs, no failure-state handling described anywhere. A competent engineer would wire this up in an n8n self-hosted instance in an afternoon with more control, better observability, and no per-seat AI tax. Skipping until there's real documentation that treats the user like an adult.”

“The HLE claim needs independent verification, but the underlying ensemble approach is architecturally sound for factual Q&A tasks. Running 339 models is expensive — pricing will be the gating factor for production use. The $10 free credit is a fair trial.”

“The direct competitors here are Zapier with OpenAI steps, Make.com, and n8n — all of which have been doing multi-step AI automations for over a year with more connectors, better error handling, and dedicated automation UX. Notion's differentiation is that the data is already there in the database, which is a real advantage for maybe 20% of use cases — the ones where your trigger and your context both live in Notion. The scenario where this breaks is the moment a user tries to do anything that requires a conditional branch or structured output parsing, at which point they're back in a Zapier tab anyway. What kills this in 12 months: Notion's core product is a notes app fighting to become a database, and every distraction into agent-land delays fixing the actual broken things (sync, performance, offline). To earn a ship, it needs to demonstrate it handles failures gracefully and show me one workflow that legitimately can't be done better elsewhere.”

“Extraordinary claims require extraordinary evidence. A 7.41 point jump on HLE via ensembling — without publishing methodology — smells like benchmark gaming. The latency of running 339 models in parallel is also a real concern for anything other than async research tasks.”

“The job-to-be-done is specific and real: 'automatically process information that lands in my Notion database without leaving the tool my team already uses.' That's a coherent single job, and Notion has a genuine distribution advantage — teams already live here, so the activation energy to automate is dramatically lower than adopting a separate workflow tool. The onboarding concern is real: building your first automation probably takes more than 2 minutes and requires understanding Notion's database model first, so non-power-users may stall. But the product has a genuine opinion — automation should live where the data lives — and that opinionated stance is the right call for a productivity suite audience. Ship with the caveat that the completeness story depends entirely on how many external integrations ship at launch.”

“The buyer is already in the room — teams paying for Notion AI at $10/member/mo just got their tier meaningfully upgraded, which is the right way to expand ARPU without a new pricing conversation. The moat is workflow lock-in: every automation a team builds in Notion is another reason not to migrate to Linear or Confluence, and that's a real switching cost that accumulates over time. The stress test is: what happens when Microsoft Copilot or Google Workspace ships equivalent automation for free to enterprise customers already paying for their suite? Notion's answer has to be 'we're faster to configure and the data model is more flexible,' which is a thin moat but a real one for the SMB segment they actually own. This isn't a transformative business move, but it's a competent defensive one that justifies the AI add-on price for another billing cycle.”

“Model ensembling is an underexplored direction in the race to reduce hallucination. If Sup AI's approach scales, it could be more durable than fine-tuning individual models — you get the wisdom of the crowd across model families, training data, and architectures simultaneously.”

“For creative work, ensemble outputs tend to regress toward the mean — you get the most-agreed-upon version of something, which is usually the least interesting version. This is a tool for factual accuracy, not creativity. I'd stick with a single strong model for writing.”